Support downsample blocks

[SuperQ]

Thanos supports downsample blocks. We keep a long history of data in downsamples so we also need to handle this in parquet.

[SuperQ]

Had a conversation with @MichaHoffmann about some of the implementation details.

We had the idea that downsample chunks should be added to an existing block, rather than new blocks like is currently done in Thanos+Prometheus TSDB.

Pros:

• It allows us to have less label index overhead.
• It allows for easy horizontal sharding, since we have all the data for a specific time range in the same block.

Cons:

• It might make different retention a little more tricky, but possibly OK if we keep the chunk files separated.

[MichelHollands]

@SuperQ I've been working on converting downsampled data to parquet in https://github.com/thanos-io/thanos-parquet-gateway/pull/35/files. It's still very much a work in progress.

The plan is to use separate chunks for the downsampled data like so:

- 0.chunks.parquet
- 0.labels.parquet
- meta.pb
- 0.chunks.5m.parquet
- 0.labels.5m.parquet
- meta.5m.pb
- 0.chunks.1h.parquet
- 0.labels.1h.parquet
- meta.1h.pb

Can you give some more information about how you were planning to proceed?

[MichelHollands]

@MichaHoffmann Do you have any more information about this?